German AI Robot Startup Raises Record Sum

Germany's NEURA Robotics raises up to $1.4 billion

NEURA Robotics from Metzingen has closed a Series C of up to $1.4 billion, led by stablecoin issuer Tether, with Amazon, NVIDIA, Qualcomm, Bosch, Schaeffler, and the European Investment Bank on board. This makes the company, founded in 2019, Europe's most highly funded humanoid manufacturer, a steep jump from its €120 million round in January 2025. The figure is an upper limit, not a booked amount, and it has increased from the approximately $1.2 billion reported by Bloomberg in March. The announcement does not mention a valuation this time (in March, it was around €4 billion). The capital comes before the robots: The flagship robot 4NE-1, listed at just under €98,000, is not scheduled to be delivered in numbers until the end of 2026. Industry-wide, over $18 billion flowed into humanoid startups in 2026, while the largest providers in 2025 collectively generated an estimated $340 million in revenue, mostly from pilots and parts rather than working fleets. Tether, which reported around $13.4 billion in profit in 2024, aims to provide the layer with its QVAC edge intelligence and WDK financial layer on which a robot completes a task, logs the result, and bills independently. → Marcus Schuler

Synthszr Take: $1.4 billion for a company whose entire industry generated $340 million last year—that's four times the combined market revenue into a single coffer. The gap between money raised and machines delivered isn't a detail here; it's the story. Matic President Mehul Nariyawala puts it bluntly: Obstacle avoidance, dodging a chair, is not yet solved; it's currently being trained. Still, I wouldn't dismiss this as a mere bubble, because the investor list reveals what's at stake: Amazon needs the hands in its warehouses, NVIDIA and Qualcomm sell the compute layer underneath, and Tether is buying into the machine that pays for itself. This is vertical integration through the European hidden champion, financed

Learning from China: Trump now also wants central filtering technology for the web

The White House and Congress are negotiating a trade-off that would reshape the open web, according to Axios. The Trump administration wants to take away states' ability to regulate AI independently (just as progressive state governments are looking to slow data center construction and hold tech companies liable). In return, a group led by Republican Marsha Blackburn is offering three federal laws: the Kids Online Safety Act, the NO FAKES Act, and a nationwide age verification. The civil rights organization FIRE, funded in part by conservative billionaire Charles Koch, warns that these packages would fundamentally change the internet. KOSA alone would allow a Trump-controlled FTC to discipline platforms like Instagram, which is regularly used by about 71 percent of US citizens. Effectively, this means the end of anonymous surfing and a lever against opposition groups. → futurism.com

Synthszr Take: Two years ago, I asked here whether the open web would become the dark web because no one moves around in it anymore. That question has been settled; now it's about controlling the pipes themselves. The real kicker is in the trade-off: AI deregulation in exchange for a central filtering infrastructure, enforced by an FTC that can be politically controlled. This is the crucial point, because if you can reach 71 percent of users through a single app, you don't need broad censorship; one pressure point is enough. In May, we wrote that Trump suddenly finds regulation sexy; here we see why, as regulation is being repurposed into a control instrument. The fact that even Koch-funded conservatives are up in arms shows that this isn't a left-versus-right issue, but a question of who's sitting at the controls. For European companies, this means concretely: don't build your reach on infrastructure that could be subject to a political filter tomorrow, but on channels with real, direct access to the customer.

Xiaomi's New Coding Agent MiMo Surpasses Claude Code in Benchmark Tests

Xiaomi's MiMo team has released MiMo Code V0.1.0 on GitHub under an MIT license, a terminal-native coding agent that, according to its own benchmarks, beats Anthropic's Claude Code: 82% vs. 79% on SWE-bench Verified, 62% vs. 55% on SWE-bench Pro, and 73% vs. 69% on Terminal Bench 2. The tool is a fork of the open-source project OpenCode, enhanced with a proprietary four-layer memory architecture (project memory, session checkpoints, scratch notes, task logs), powered by SQLite full-text search. The real trick: A separate checkpoint-writer sub-agent continuously logs in the background while the main agent builds, so that when the context window is full, the working environment is reconstructed from structured checkpoints instead of being lost. In a double-blind A/B test with 576 developers in 474 real repositories, both systems were at 50/50 for tasks under 200 steps, but beyond 200 steps, MiMo's win rate rose above 65%. Notably, Xiaomi compares itself exclusively with Claude Code, not with OpenAI's Codex (GPT-5.5 scores about nine points higher on Terminal-Bench 2.0 at 82.2%) or Google's Gemini CLI. Additionally, there is temporary free access to the multimodal flagship MiMo-V2.5 with a million-token context window, no registration required. → VentureBeat

Synthszr Take: The interesting number isn't in the benchmark table, but in the A/B test: a tie on short tasks, then suddenly a 65% win rate beyond 200 steps. This is precisely the thesis Xiaomi is selling here, and it could be right: the harness around the model is becoming as important as the model itself. A five-point lead on SWE-bench Pro comes purely from the agent system, with an identical underlying model. This aligns with what we already saw in the Cursor-Anthropic debate in March: competition is shifting from the raw model to orchestration, memory, and state management. Still, caution is advised—these are self-reported numbers, and the consistent absence of Codex and Gemini in the comparison says more about PR selection than about the true top performers. For anyone building a coding stack: MiMo Code is MIT-licensed, installed with a curl command, and thus a serious open-source candidate for regulated or lock-in-sensitive workloads. Trying it out costs an afternoon, and the memory approach is the right lever in the right place.

Companies Spend Up to $7,500 Per Employee Per Month on AI

A Nvidia manager recently said that compute costs now exceed the salaries of his people. Mercor's CEO followed up last week: his startup pays more for tokens for internal agents than for its workforce. However, the new Ramp AI Index puts this into perspective. The top 1 percent of US companies, dubbed “AI-pilled” by Ramp, spend $7,500 per employee per month, which is well below the roughly $16,000 that an average software engineer costs monthly. The top 10 percent spend about $611, while the median is a meager $11.38, the price of a single enterprise seat. Despite all the cost debates, spending at AI-pilled companies increased by 14.1 percent per capita last month. The front-runners mix several frontier models with cheaper open-source options. → Techpresso

Synthszr Take: $7,500 a month sounds brutal until you compare it to a $16,000 engineer salary. Then it's a calculation that can be decided tomorrow morning, not after the next strategy offsite. The real signal is in the behavior of the leading group: they jump between frontier models and cheap open-source variants because they understand that model choice is fluid and compute discipline is the real lever. In May, we were still writing about token shock; now we see who is processing it productively and who is just staring at the bill. Those who aggregate token consumption at the use-case level instead of waiting for a black-box cloud bill are running a TCO of around €2,730 per engineer per year with an ROI of eight to fifteen times. The 14.1 percent growth isn't an alarm; it's Jevons' paradox in its purest form: the cheaper the inference, the more you consume. The dividing line is between companies that see every call and companies that hope it will all work out somehow.

long before a single 4NE-1 pays for itself. For Europe, the real question isn't whether NEURA deserves the valuation, but whether Metzingen can maintain its velocity when the first delivery date at the end of 2026 becomes a test. Anyone who raises money before the product has exactly one job: deliver before the momentum shifts.

Data Centers: China Bets on Large Reactors

While the US has only brought two new reactors online in years (Units 3 and 4 at the Vogtle plant in Georgia), China is building at a record pace: six new reactors were started in 2025, with two more in the first five months of 2026. By 2030, the country will surpass both the US and the EU in installed nuclear capacity. The speed advantage is enormous: a Chinese reactor is built in five to seven years, the global average is nine, and the last US reactors took about 15 years. The lever is standardization, uniform project management, and construction in series of six or more units for economies of scale. Meanwhile, the West is betting on small reactors: Antares from California just reached criticality with its Mark-0 in the US Department of Energy's pilot program, which aims to have three test reactors running by July 4, 2026. Big Tech like Google is investing in such mini-reactors to power data centers, but large plants deliver electricity more cheaply per kilowatt-hour. → The Download from MIT Technology Review

Synthszr Take: China is winning here with the most boring idea in the world: build the same thing six times instead of starting from scratch each time. Standardization beats ingenuity, and that's precisely what the West finds so difficult, because every project there becomes a one-off, with its own approval process (costs in the billions, amortization in decades). Velocity is the real story. Antares just reached criticality with the Mark-0, but the plant has neither heat dissipation nor power conversion; the first kilowatt-hours won't come until the end of 2027 at the earliest. This is an honest prototype and not a power supplier, while in China, the Linglong-1 is going online this year, and eight large reactors are being built simultaneously. The bet on small reactors can pay off if they truly come off a factory line and aren't reinvented with each new plant. Until then, the sober reality is: if you want to put electrons on the grid, you build what you already know how to build, and you do it in series.

Drones are now also killing completely autonomously on the front lines

A high-ranking representative of the Ukrainian defense industry has confirmed what was previously only suspected: fully autonomous drones with no human control whatsoever have killed soldiers on the battlefield. Drone builder Alexander Kokhanovskyy described to New Scientist a test conducted about two years ago with ten AI-controlled quadcopters near Bakhmut and Chasiv Yar. The devices flew 3 to 5 kilometers to the front line in about ten minutes and then switched on “Terminator mode,” in which an AI model independently seeks and attacks targets. There was no radio link, no video feed, no way to intervene: 'Everything it sees is killed.' Only afterward did human pilots send drones into the area to check the results: a few soldiers and a truck. Ukraine officially still prohibits AI in the final step of targeting, but according to Kokhanovskyy, it is already negotiating looser rules. → The Download from MIT Technology Review

Synthszr Take: The 'human-in-the-loop' doctrine, which all the world's militaries have invoked for years, quietly died with this one test. Ten drones, no human in the loop, one button press, and the system then decides on every target in the area on its own. Technically, this isn't a leap, but a combination of components anyone can assemble: a quadcopter for a few hundred dollars, an image recognition model, and some control logic. That's precisely what makes it so hard to contain. In 2023, the word was that AI drones only attacked tanks, not people; now the government is openly talking about loosening its own guardrails because the enemy doesn't abide by them. Anyone who thinks this will remain a special Ukrainian case has slept through the last two years.

Instagram now lets you tell the algorithm what you want

Instagram is opening up its main feed to user control. With the 'Your Algorithm' feature, the app shows the topics it thinks you're interested in and lets you change them across all key areas. So far, it's just topics, but according to Instagram head Adam Mosseri, requests for people, moods, and content types are coming. The feature was already running for the Reels feed and the Explore page. Mosseri gets philosophical: 'This is the start of something bigger than a feature,' saying it's in the business's interest to give people more agency over the products where they spend so much time. The crucial technical point: only large language models can describe content clusters in understandable language, so the system can show you what it thinks about you, and you can tell it what you really want. → Techpresso

Synthszr Take: Mosseri has delivered an honest diagnosis, and that's rare enough. The unease with social media doesn't come from the content, but from the feeling that the experience happens *to* you, rather than you shaping it. For years, the feed was a black box that learned from your taps without you being able to talk back. Now, LLMs are making the ranking logic readable, and that shifts something fundamental: giving an actor at the edge more decision-making power means relinquishing control. Mattes' point from Code Crash applies doubly here: visibility is the lever, not the model. It will be interesting to see if Meta designs this new transparency in a truly user-centric way, or if the topic tags will just end up being a prettier control panel for the same time-on-site machine. We wrote at the end of February that personalization is sometimes crap; finally giving it a hand on the wheel is the right move, as long as the wheel is actually connected to the steering.

Skepticism Resides in the Rich West, Optimism in Emerging Markets

Salesforce and YouGov surveyed 1,500 office workers across four continents, and the result upends typical expectations. Americans are 43 percent more likely to describe themselves as AI skeptics than the global average; over half of US employees classify themselves this way. In India, both trust and daily usage are over 80 percent, while in the US, both figures hover around the 50 percent mark. In emerging economies, growth is consistently expected, with 90 percent anticipating a career boost from the technology. The finding: where the models are built, the distance grows, and where the need to catch up is greatest, the appetite grows. → AI Secret

Synthszr Take: These numbers align suspiciously well with what we know from Europe. Germany's corporate adoption rate is 26 percent, the EU average is 20, while US adoption ranks above 85 percent (Eurostat 2025, Stanford HAI 2026). And yet, US employees are the loudest doubters, while India just gets on with it, with over 80 percent daily usage. The explanation is uncomfortable: those who already have a lot protect their identity, their hierarchy, and their position; those who want to catch up have nothing to lose and everything to gain. Skepticism here isn't a knowledge problem, it's a privilege of prosperity, and that eats up time. In May, we wrote that Zuckerberg, Musk, and others are shattering public trust in AI; this survey shows who pays the price for that. Anyone who wants to bring their workforce along in 2026 needs to cure the fear of their roles being redesigned, not just their knowledge of the tools—and that can start in the next team meeting, not after the next strategy offsite.

GEO: Even Shopify is doing listicles

Shopify has published at least 60 listicles on its blog, from '10 Best Ecommerce Platforms for Small Business in 2026' to 'Best Ecommerce Software 2026'. The competitors in the lower ranks change, but the top spot always belongs to Shopify. The audience isn't people, but chatbots: anyone who asks ChatGPT for the best way to start an online store gets told Shopify, with Shopify's own rankings as the source. Figma has at least six such lists, ClickUp almost 300. Tom Critchlow of Raptive reports that a Google search with an AI answer leaves other sites with only half the traffic, and that ChatGPT-links are below 0.5 percent. The SEO industry is christening the new discipline GEO, Generative Engine Optimization, and is feeding the models with self-written best-of lists that no human ever reads. → The Download from MIT Technology Review

Synthszr Take: The problem isn't the audacity, but the naivety of the models. A chatbot that can't distinguish an 'about us' list from Shopify from an independent test is still unready to be a gatekeeper. This is precisely where the entire industry's thinking is flawed. Anyone who thinks they can permanently write their way to the top with 300 of their own listicles like ClickUp is flogging a dead horse, because studies have long shown that mentions on Reddit, in trade publications, and on review platforms correlate more strongly with AI visibility than anything a brand says about itself (Ahrefs measures 0.664 versus 0.218). In B2B, only one in eight SaaS brands shows up at all when buyers query their category; the other 88 percent simply don't exist in the model. The self-congratulation only works as long as OpenAI and Google don't seal off their models, and that window is closing faster than the 'sloptimisers' can type their lists. The productive question, therefore, is not 'How do I rank myself higher?' but 'What word in my category has no one occupied yet?'. This exercise can be started tomorrow morning, long before the next listicle factory gets running.